---
title: OpenAI Extends Prompt Cache Retention to 24 Hours
description: OpenAI extended prompt cache retention from a few minutes to up to 24 hours. What that changes for agents and pipelines with repeated calls.
date: 2026-08-12T00:00:00.000Z
category: ai-productivity
tags: tokens, gpt, prompt-caching, openai
---

## Quick Answer

OpenAI extended prompt cache retention from a short, few-minute default to up to 24 hours. Workflows that make repeated calls sharing a prompt prefix, like an agent looping over the same system instructions, can now rely on a cache hit hours later instead of only minutes later. Re-check any workaround you built for the old short window.

## What's actually changing

Third-party cost-analysis coverage (effloow.com), dated 2026-05-29, reports that OpenAI moved the default prompt-cache lifetime from a few minutes to up to 24 hours. That is a meaningfully longer window for anything that reuses the same prompt prefix across separate calls that are not tightly clustered in time.

This is paraphrased from measured third-party analysis, not confirmed against OpenAI's own primary documentation directly, and is stated honestly as such. The mechanism it changes is retention duration specifically, distinct from the cache-write premium and breakpoint requirements covered separately, and [OpenAI's own prompt caching guide](https://platform.openai.com/docs/guides/prompt-caching) documents an up-to-24-hour extended retention window using the same figure for eligible models.

## Structural Comparison Matrix

| Operational Aspect            | Before                                    | After                              |
| :---------------------------- | :---------------------------------------- | :--------------------------------- |
| **Default cache lifetime**    | A few minutes                             | Up to 24 hours                     |
| **Repeated-call workflows**   | Needed calls within minutes to hit cache  | Can span hours and still hit cache |
| **Workarounds for short TTL** | Common (keep-alive pings, tight batching) | Likely unnecessary now             |

## Fix it: stop working around a window that no longer exists

If your pipeline previously batched calls tightly, or sent a periodic no-op request, specifically to keep a cache entry alive before it expired, that workaround is likely solving a problem that no longer exists at this scale. A 24-hour window covers most batch jobs, scheduled agent runs, and even a full business day of interactive use without needing an artificial keep-alive.

The more useful change is upstream: this is a good time to restructure prompts so the reusable part, a system instruction, a long reference document, a tool schema, sits at a stable prefix marked as a cache breakpoint (see [GPT-5.6 Ends Free Prompt-Cache Writes](/ai-productivity/gpt-5-6-prompt-cache-write-premium/) for how breakpoints work). A longer retention window makes that investment pay off across a wider span of real usage than it did before.

<Callout type="tip" title="Model the reuse window, not just the write cost">
  A longer retention window only pays off if your workload actually reuses the
  cache inside it. The [LLM Pricing Calculator](/tools/llm-pricing-calculator/)
  lets you compare the cache-write cost against accumulated read savings for
  your own call pattern, rather than assuming a 24-hour window automatically
  makes caching worthwhile.
</Callout>

## Confirmed version

Sourced from effloow.com's cost-analysis coverage, dated 2026-05-29, not OpenAI's own primary documentation directly. Browse more coverage in the [AI Productivity](/ai-productivity) archive, or start from [The 2026 LLM Token & Pricing Reset](/ai-productivity/2026-llm-token-pricing-reset/) hub.